Video models
FLUX 3 Max: fal's post-trained video model makes 10 seconds in 12
fal Research has post-trained Black Forest Labs' FLUX 3 video model into FLUX 3 Max, which it says ties for first in image-to-video quality while generating a 10-second clip in about 12 seconds. It is live on fal's API.
fal Research has launched FLUX 3 Max, a video generation model post-trained from Black Forest Labs’ FLUX 3 in collaboration with the lab. It went live on 15 October, the day fal’s launch pricing began. The model is available on fal for text-to-video and image-to-video, with a keyframes endpoint linked from the announcement.
What it is
FLUX 3 is the first FLUX model to generate video. Black Forest Labs built the base: one model trained jointly across image, video, audio and action prediction, producing clips of up to 20 seconds with sound rendered in the same pass.
fal says Black Forest Labs supplied the foundation model, while fal Research added post-training, inference optimisation and systems engineering. fal describes it as the same approach it used for H3 Max, an earlier model post-trained from MiniMax’s open weights. fal adds that it will keep post-training FLUX 3 Max.
The numbers
All of these are fal’s own figures, and the quality ranking is described as preliminary.
- Quality: 1,139 Elo in Artificial Analysis’s preliminary image-to-video arena, which fal reports as tied for first, level with Seedance 2.5.
- Improvement: a gain of roughly 50 Elo points over the original FLUX 3 (fal gives 47 on the model page and 50 on the product page, 1,139 against 1,089).
- Speed: a 10-second video in 12.45 seconds, against 277.80 seconds for Seedance 2.5 in fal’s benchmark. The blog calls that 26.5 times faster; the product page rounds to about 22 times, and the model page says about 11 seconds and more than 25 times. The exact multiple varies by page, but all agree it is more than an order of magnitude.
What you can control
According to the model page, clips run from 5 to 20 seconds in one-second steps, at 720p by default with optional 1080p. Audio can be generated alongside the video or switched off. Aspect ratios include 16:9, portrait, square and wider cinematic shapes. A low or high reasoning setting trades time for quality. Image-to-video accepts an optional ending image.
Price and access
FLUX 3 Max is served through fal’s API. The model page lists promotional prices of $0.0425 per second at 720p and $0.0725 at 1080p, 75% off, from 15 October to 7 November. After that the page says 720p is $0.17 per second and 1080p $0.29. fal’s product page also says signed-in accounts can generate five FLUX 3 Max videos a day for free. The pages do not mention downloadable weights for Max; Black Forest Labs has said an open-weight FLUX 3 Dev is planned.
Limits
The ranking comes from a preliminary arena and from fal’s own benchmark, not an independent speed test. Clips are capped at 20 seconds, and the model produces video, not a playable, stateful world.
Our take
Generation speed changes what a video model is good for. At about 12 seconds for a 10-second clip, trying a dozen takes becomes cheaper than agonising over one prompt, which suits cutscenes, previews and generative interactive film, where a branch might be rendered just before the player reaches it.
It is still a clip generator, though. It does not hold rules, inventory or a character’s memory between shots; that job sits with the game, as we argue in how AI world models work. Faster video makes the visual layer of an interactive story cheaper to produce, but the story logic still has to live somewhere else.